---

name: lora_parameter_budget description: Compute LoRA trainable parameter budgets and reduction ratios for adapted Transformer projections.

LoRA Parameter Budget

Use this skill to estimate how many task-specific parameters LoRA trains compared with full fine-tuning.

Inputs

Projection dimensions or d_model, rank r, adapted matrix count, layer count, and optional full-parameter baseline.

Outputs

Trainable LoRA count, dense-update count for the same matrices, and reduction ratios.

Workflow

For square Transformer projections use the paper formula 2 * L_hat * d_model * r, where L_hat counts adapted matrices. Compare against full dense updates L_hat * d_model * d_model.

Validation

Run included tests for formula correctness and rank monotonicity.

Limitations

This is an accounting skill; it does not measure wall-clock training throughput.